Qwen/Qwen3.5-9B

Tinker's LoRA fine-tuning matches full fine-tuning performance on small datasets and reinforcement learning, but underperforms on larger supervised learning datasets, according to a technical primer f…